Herburger (1997) and Ahn & Sauerland (2017) propose that the quantifier is base-generated inside the DP, and moved out of the DP. While details differ, they both argue that the displaced determiner takes a focus-generated restrictor argument, thus deriving the non-conservative reading. We discuss the two approaches below.

Lastly, it has also been argued that the counterexamples to the Conservativity Universal are not indeed counterexamples, because expressions such as only, many and few are not in fact determiners. For example, the non-conservative property of only is not taken to be a counterexample to the Conservativity Universal because only is analyzed as an adverbial (von Fintel & Keenan 2018; von Fintel 1997). Semantic analyses of only treat it as taking the full proposition as an argument, rather than just taking the content of the NP. The question that follows is whether the non-conservative floating proportional quantifiers in Korean, and more broadly all non-conservative proportional quantifiers discussed in Ahn & Sauerland 2017, are adverbial rather than adnominal. For example, it is possible that the proportional measure 60% in SNU hired 60% students is an adverbial that attaches in the VP and quantifies over hiring events.

In such an analysis, 60% would behave similarly to an adverbial like mostly in SNU hired mostly students. The main arguments that Ahn & Sauerland (2017) present against this view involve showing the constituency of the DP that involves the determiner and the noun. First, for languages like English, a measure phrase like 75% cannot be separated from the restrictor noun like mostly can.

Second, preposing 75% women is possible, which is not predicted if 75% is an adverbial because it would not form a constituent with women.

However, these criteria are not applicable to Korean given that the non-conservative construals always involve floating quantifiers, which by definition are separated from the host noun. Ahn & Sauerland (2017) do not provide any Korean-specific diagnostic to determine whether the floating quantifiers are indeed adnominal or not, thus leaving this option open for investigation.

Moreover, there is a related work in Korean that argues for an adverbial analysis of a non-conservative floating quantifier. Park (2007) shows that the quantifier taypwupwun (‘most/mostly’) allows a non-conservative reading when floated, and proposes an adverbial account. More specifically, she argues that taypwupwun should be analyzed as an adverbial quantifier that quantifies over events. For example, (15) is analyzed as (16) (modified and translated from Korean).

Given that there are no diagnostics provided to determine whether the floating quantifiers involved in non-conservative construals are adnominal or adverbial, and that the only other account of a non-conservative proportional quantifier analyzes it as an adverbial element, it seems crucial to confirm whether the proportional quantifier like 70% in Korean is indeed base-generated in the nominal phrase as Ahn & Sauerland 2017 argue.

What we see is that in both of the adnominal accounts, the quantifier is argued to move out of the DP. In Herburger 1997, it is done through D-raising, and in Ahn & Sauerland 2017, it is done through QR. In either case, movement is a crucial element of the adnominal accounts, because only after movement the quantifier combines with a focus-generated restrictor argument.

Korean is argued to provide overt evidence for movement in Ahn & Sauerland 2017. Korean has been argued to mark scope overtly (Beck & Kim 1997). In Korean, the non-conservative construal involves a floating quantifier structure unlike the conservative construal. This is shown in (17), where the proportional quantifier 20% appears inside the nominative-marked subject for the conservative reading in (17a), and outside the nominative-marked subject for the non-conservative reading in (17b).

Ahn & Sauerland (2017) argue that while the QR of NP₁ in other languages is covert, it is reflected in the syntax in Korean, thus supporting the QR-based analysis.

However, the assumption that a floating quantifier derives from a DP-internal position is not readily supported in the literature. In fact, what we observe is that the discussion is quite divided. The (non-)local dependency between a noun and a floating quantifier has been widely debated in the syntax literature. One of the main debates are formed around the question of whether the floating quantifier is adnominally derived, meaning that they are base-generated inside the DP and stranded from the NP, or adverbial, meaning that they are base-generated outside the DP and inside the VP. Ko (2014) argues that some languages have both adnominal and adverbial floating quantifiers, and argues that Korean is one of those languages that have both adnominal and adverbial types (see also Ishii 1998, Kang 2002, Fitzpatrick 2006, among others).2 Ko proposes that adnominal floating quantifiers are externally merged to its host NP in the base structure, as schematized in (17) (see Kuroda 1983; Sportiche 1988; Ueda 1990, a.o.). Adverbial floating quantifiers, on the other hand, contain a pro linked to the host NP and are merged outside a verbal structure, as described in (18) (see Doetjes 1997, Kobuchi-Philip 2003; 2007; Fitzpatrick 2006).

Based on a range of semantic and syntactic diagnostics, Ko (2014: Chapter 3) argues that bare numeral quantifiers without focus belong to the adnominal type in Korean and that floating quantifiers with exhaustivity (e.g. Case-marked numerals, exhaustive-marked numerals, motwu ‘all’, NPI amwuto ‘anyone’) belong to the adverbial type. As far as the authors know, there is no syntactic work that investigates whether proportional quantifiers such as 70% is adnominal or adverbial underlyingly.

In order to determine whether Korean floating quantifiers in non-conesrvative construals involve adverbials or adnominals, we designed two experimental tests based on different diagnostics. In the first study, we test whether the host noun can be definite-marked, based on our reasoning discussed above that adnominal quantifiers would not allow definite host nouns. In the second study, we test whether non-conservative reading is possible when a floating quantifier is separated from the host noun by a low adverb, which is a diagnostic for identifying adnominal floating quantifiers presented in Ko 2014. Because differentiating between conservative and non-conservative construals requires identifying subtle differences in interpretation, we tested our target sentences against a larger number of participants in a systematically controlled experimental setting.

Korean has a three-way distinction in demonstratives (Sohn 1994): i is a proximal demonstrative used to pick out entities close to the speaker, while ce is a distal demonstrative used to pick out entities far from both the speaker and the hearer. The third demonstrative ku is also traditionally analyzed as a distal demonstrative that picks out an entity present in the context, but more recent studies show that ku is an anaphoric marker (Ahn 2017; Ahn & Davidson 2018; Cho 2016): it requires the addressee to be familiar with the intended referent. The notion of familiarity that we use here requires the entity to have been previously mentioned in the discourse, which we call strong familiarity following Roberts 2003. This term contrasts with the broader term weak familiarity, which also includes entities that have not been mentioned previously but nevertheless familiar to the addressee (Roberts 2003). For example, in (21), the sentence would be unacceptable if the book in question has not been previously mentioned and established in the discourse. Thus, the insertion of ku allows us to restrict the host noun to an anaphoric, definite reading.

There is another phenomenon that interacts with definiteness in Korean, which is number marking. Korean is a generalized classifier language that makes use of classifiers to individuate nouns into countable units. However, unlike other classifier languages like Mandarin and Japanese, it has a productive plural marker tul that can combine with nouns to imply plurality. Studies have shown that the plural marker is obligatory when associated with definite readings (Kim 2005; Ahn & Snedeker 2021; Kwon & Zribi-Hertz 2004). For example, plural marking on chaeg (‘book’) is obligatory if it is anaphorically referring to a previously introduced book in (22).

Summarizing the two points above, there are two grammatical properties that indirectly mark definiteness: the anaphoric demonstrative ku and plural marking. These two mechanisms can be combined to form a minimal pair of sentences that only differ on whether the noun is interpreted as definite or not. Consider the two sequences below:

The first sentence introduces a discourse referent of two-hundred women who received the company advertisement. The only difference between (23a) and (23b) is that while the former contains a demonstrative description with a plural marker ku yeca-tul (‘those women’), the latter only involves a bare noun yeca (‘woman’). The continuation in (23a) is restricted to a definite, anaphoric reading, where ku yeca-tul refers to the women mentioned in (23): (23a) asserts that forty of those women attended the show. On the other hand, (23b) can only have an indefinite reading that there exist forty women who attended the show, with no implication that the forty women are associated with the discourse referent introduced in (23). Even though the setup of the example pragmatically makes an anaphoric continuation more likely, a definite reading of the bare noun in (23b) is impossible because plural marking is absent: if interpreted as a definite, the absence of the plural marker in (23b) would suggest that there is only one salient woman, but that reading is incompatible with the cardinality of forty that follows.

Note that it is possible to only have the plural-marked noun without the demonstrative as in (24), but in that case, the meaning is ambiguous between a definite and an indefinite construal. In order to avoid this ambiguity, we focused only on sentences like (23a) and (23b).

By replacing the absolute cardinal forty to a proportional measure like 40%, we are able to construct a minimal pair of sentences that only differs on whether a definite reading of the host noun is available. To test whether definiteness of the host noun affects the availability of the non-conservative reading as the adnominal accounts predict, we created an experimental stimuli with a 2 × 2 design between two factors: a) whether the host noun was marked as definite with both a demonstrative and plural marking ([±def]), and b) whether the proportion was compatible with a conservative or the non-conservative reading ([±con]) given a context. This created a total of four conditions: [+def,+con], [+def,–con], [–def,+con], and [–def,–con].

In the study, we presented the target sentences with a context that allowed the participant to determine whether the given cardinality in the target sentence was a proportion from the total set denoted by the host noun (a conservative reading) or from the total set denoted by the predicate in the VP (a non-conservative reading). For example, for each of the four trials, participants first read the context in (25). For convenience, we provide the English translation of the Korean sentences.

Below the context, one of four target sentences was presented to the participant. All four target sentences started with the sentence in (26), followed by one of the four sentences in (27). In the [+def] conditions, the host noun appeared with a demonstrative description and plural marking, while in [–def] conditions, a bare noun was used. In the [+con] conditions, the proportion used was 40%, while in [–con], it was 80%. This is because the conservative reading in (27) would compare the number of women who attended the show to the number of women who received advertisement (80/200; 40%), while the non-conservative reading would compare the same number to the total number of people who attended the show (80/100; 80%).

After reading one of the target sentences, the participant was asked to judge whether the target sentence was true or false given the context in (25). If there is a restriction against definite-marked host noun, we would predict (27a) to be judged to be true and (27b) to be judged as false. The hypothesis does not make a specific prediction for the [–def] conditions in (27c) and (27d): both conservative and non-conservative construals should be available in principle.

This prediction aligns with the authors’ judgments. Given the context in (25), (27a) is judged to be true, and (27b) is judged to be false, as both sentences only allow a conservative reading. With the non-definite marked host nouns, the authors judge (27d) to be true: it readily allows the non-conservative reading where 80% of those who came were women. It is unclear to the authors whether (27c) readily allows the conservative reading as well. While the conservative reading should be available in principle, it is not salient, and some sort of a contrast is required to allow the conservative reading.

Because the judgments are subtle, some noise and interspeaker variation are unavoidable in the data. However, we looked for statistically significant contrasts across conditions and made conclusions based on such contrasts.

Sixty-one native, monolingual speakers of Korean participated in the study. They were recruited by advertising on the campus of Seoul National University in Seoul, Korea, with flyers and on social media. The participants were asked to complete an online survey created through PCIbex (Zehr & Schwarz 2018). After Study 1 on definiteness was complete, they were given an option to proceed to Study 2 on adverbs. Two participants were removed from the analysis because they participated in the survey twice.

The study was conducted online at the participants’ respective homes. Three background questions were asked at the beginning of the survey: age (in decades), dialect of Korean used (Chuncheon, Gyeongsang, Jeju, Jeolla, and Seoul/Kyungki), and English exposure (native, fluent, some familiarity, and none). After the background survey, they were instructed to read the stimuli consisting of the context in (25) followed by the target in (26) with one of the four target variations in (27), and respond whether the target sentence in (27) is true or false.

There were a total of four test trials and four filler trials. The conditions were organized in a Latin-square design for a within-subject study, so each participant saw one of the four conditions once. A screenshot of a trial is shown in Figure 1.

The participants’ responses in each trial were coded as ‘1’ for ‘true’ and ‘0’ for ‘false’. With the Latin-Square design and four groups, there were 11 to 20 responses per trial, with a total of 58 responses per each of the four conditions. A generalized linear mixed-effects model (GLMM) was used for the analysis, with construal and definiteness as fixed effects and group, participant, and item as random effects. The average proportion of ‘true’ responses by Construal (conservative vs. non-conservative) and Definiteness of host noun (definite vs. non-definite) is plotted with standard errors in Figure 2. There was a main effect of construal: participants were significantly more likely to respond ‘true’ in the [+con] condition, i.e. when the proportional measure was compatible with the conservative construal (Z = 2.006, p = 0.04). There was also a significant interaction between Construal and Definiteness (Z = 4.767, p < 0.001). In the conservative conditions, participants were significantly more likely to respond ‘true’ when the host noun was definite (89% [+def] vs. 59% [–def]; Z = 3.352, p < 0.001), while in the non-conservative conditions, they were significantly less likely to respond ‘true’ when the host noun was definite (40% [–def] vs. 10% [+def]; Z = –3.385, p < 0.001).

First, we note that the acceptance rates for the two non-definite conditions are at around 50%. This is very different from the rates in the definite conditions, where there is a clear indication of accepting the conservative construal and rejecting the non-conservative construal. Because participants only saw one test item in each condition, the near-50% rate does not suggest that participants were responding at random. Instead, this suggests that 40% of the participants (24/58) responded ‘true’ in the [–def,–con] condition, while 60% responded ‘false’ in the same condition. As an anonymous reviewer pointed out, participants are biased towards true construals of ambiguous sentences when they are available. Assuming this is the case, we would expect the acceptance rate of the [–def,–con] condition to be at-ceiling if the reading is available. The near-50% acceptance rate thus does not definitively demonstrative the availability of the non-conservative construal with floating quantifiers for all Korean speakers. In order to test whether the results are actually indicative of interspeaker variation or due to general confusion in the [–con] conditions, a follow-up with more items per condition, or with a third ‘I don’t know’ option as the reviewer suggested, would be necessary. The weaker conclusion we can draw from this set of data for now is that the non-definite-marked noun is not as systematically rejected as the definite-marked noun in the non-conservative context.

As for the effect of definiteness, the adnominal and the adverbial accounts made different predictions. The adnominal account predicted the [+def,–con] condition to be rejected, while [–def,–con] condition to be accepted. This prediction is borne out: participants reliably rejected [+def,–con] condition (acceptance rate 10%), while they accepted [–def,–con] at a significantly higher rate at 40%. This difference was not predicted in the adverbial account, because adverbial quantifiers do not restrict host noun type.

We further find that the conservative reading is available in sentences containing a non-definite bare noun, with participants judging the sentence to be true 59% of the time. There are a couple of ways in which the conservative reading can be derived under the adnominal accounts. First, because focus generates the content of the restriction, an interpretation equivalent to a conservative construal is possible if the VP is focused. It could be argued that the participants took the target sentence in this condition to be VP-focused, resulting in the conservative reading. The possibility of deriving a conservative reading by focusing VP is discussed in both Herburger 1997 and Ahn & Sauerland 2017. Second, the floating quantifier could have been treated as a case-marked floating quantifier. This is possible because in Korean, case markers can be dropped optionally. We discuss the ambiguity between conservative and non-conservative construals further in Section 6.2.

In summary, the results of Study 1 show that when the host noun is definite, a floating quantifier cannot receive a non-conservative reading in Korean. This is predicted if the non-conservative construal results from an adnominally-derived quantifier, but not if it results from an adverbial quantifier. Thus, our results support the adnominal account of non-conservative quantifiers.

It has been extensively argued by Ko (2007; 2014) that sub-extraction out of syntactic edges is severely restricted and underlying constituency makes a crucial difference in sub-extraction. In particular, elements merged as a constituent on syntactic edges of a cyclic domain cannot be separated by their domain-mate. This ordering restriction is called the Edge Generalization (EG), stated in (28). The EG results from a conspiracy of cyclic linearization (Fox & Pesetsky 2005) and locality constraints on movement. If two elements, X and Y are externally merged on the syntactic edge, they cannot move around within the Spell-out domain (under Chomsky 2001 for probe-goal Search), so that X and Y may be followed or preceded by their domain-mate Z, but they cannot be intervened by Z. This ordering restriction holds throughout the derivation due to cyclic linearization, which maintains the monotonicity of orderings after Spell-out.

Ko (2014) presents supporting evidence for the EG, based on sub-extraction and (re)-ordering patterns out of syntactic edges in predicational domains in Korean and Japanese. Some well-known restrictions on the distribution of the subject and an adnominal quantifier can be understood as an instance of the EG. The subject and a bare numeral in Korean cannot be separated by their vP-domain mate such as the object, as illustrated by the contrast between (29a) and (29b). The bare noun also cannot be separated from the subject by a vP-internal (low) adverbial as shown in (30a). The subject and the adnominal quantifier are merged in the specifier of a Spell-out domain vP, and thus they cannot be separated by a vP-internal element within vP, and this ordering restriction is preserved under cyclic linearization. A vP-external high adverbial, on the other hand, may intervene between the subject and its adnominal Q, as shown in (30b). This is because the subject can undergo scrambling to the left of the high adverbial outside vP without causing an ordering conflict.3

Adverbial floating quantifiers, on the other hand, behave differently from adnominal floating quantifiers. Since adverbial quantifiers are merged outside vP, separately from the host noun, a vP-internal element may intervene between the subject and its associate adverbial quantifier. Ko (2014) shows that a case-marked numeral FQ (floating quantifier) and an exhaustive-marked numeral FQ exhibit this behavior. When the numeral FQ carries a case-marker or an exhaustive marker man (‘only’), the quantifiers allow separation from the host noun by a domain-mate as in (31) and by a low adverbial in (32). This sharply contrasts with what we observe with bare numeral quantifiers in (29b) and (30a), which disallow such separation by a vP-domain-mate.

Building on these arguments, we test how the presence of a low and a high adverbial affects the availability of the non-conservative reading.

As discussed in Section 4.1, an adnominal quantifier in Korean differs from an adverbial quantifier in that low adverbs such as ‘intentionally’ and ‘quickly’ cannot intervene between the host noun and the quantifier. If the floating quantifier involved in the non-conservative construal is an adnominal quantifier, we predict that the non-conservative reading would not be available when a low adverb intervenes. If the floating quantifier is adverbial, on the other hand, we would predict the availability of the non-conservative reading to be not affected by the type of adverbs that intervenes.

The authors’ judgments align with the prediction of the adnominal approaches. For example, in (33a), where a high adverb pwunmyenghi (‘certainly’) intervenes the host noun and the proportional measure noun, both conservative and non-conservative readings are available. However, when a low adverb such as capalcekulo (‘voluntarily’) or sengsilhi (‘diligently’) intervenes, the non-conservative reading is no longer available.

In order to confirm that this contrast is valid and reliable across a larger number of speakers, we designed an experimental study that tests the predictions. We presented four different sentences similar to (33) with a context that was compatible with either a conservative or a non-conservative reading, creating a 2 × 2 design between two factors: a) whether the intervening adverbial is high or low ([high], [low]), and b) whether the context provided was compatible with a conservative or a non-conservative reading ([±con]). Because the prediction in Ko 2014 is based on grammaticality, we asked the participants to not only judge whether the sentence is true or false, but also how natural the sentence is out of a six-point Likert scale.

The four target conditions [high,+con], [high,–con], [low,+con], and [low,–con] were designed as follows. In the [high] conditions, a high adverbial like surprisingly and certainly appeared in both the context and the target sentence. In the [low] conditions, a low adverbial like intentionally and diligently appeared in both the context and the target sentence. The adverbials were added to the context prose because the target sentences sounded odd when the adverbials suddenly appeared out-of-the-blue without contextual support. The target sentences only differed in the choice of the adverb, as shown in (34), with the authors’ judgments presented in the gloss.

The [±con] factor was manipulated in the context prose. For example, in the [+con] conditions, the participants saw the prose in (35) followed by (35a). Because both of the target sentences contained 20%, the numbers in (35a) is only compatible with the conservative reading of the quantifier, which compares the number of doctors who participated against the total number of doctors. In the [–con] conditions, on the other hand, the participants saw the prose in (35) followed by (35b), with the numbers compatible only with a non-conservative reading, which compares the number of doctors who participated against the total number of people who participated.

The conditions that would tease apart the two approaches are the [low] conditions: the adverbial approach does not predict the acceptance rate for [low] conditions to differ based on the given context [±con], while the adnominal approach predicts the acceptance rate to be much lower in the [–con] condition than in the [+con] condition. The two approaches do not make specific predictions for [high] conditions because high adverbials may appear with either the adnominal or the adverbial quantifier. Thus, we predict the sentences in [high] conditions to be accepted at a high rate unless other factors interfere with the interpretation of high adverbs.

The participants who finished Study 1 were asked to follow a link that led them to Study 2. Sixty out of 61 participants continued onto Study 2. The general methodology of Study 2 was the same as that of Study 1. Because the prompts for this study were more involved, with many numbers appearing in the context, we added Venn Diagrams to help them understand the relations. For example, Figure 3 was shown on the screen with the [high,–con] condition of the trial shown in (35b).

Below the context paragraph and the Venn Diagram, the target sentence as in (34) was presented in the middle of the screen. Based on the condition the trial was in, the target sentence contained a high intervening adverb as in (34a) or a low intervening adverb as in (34b). After reading the target sentence, the participants were asked to judge whether the target sentence is true or false based on the presented context. After they responded, a six-point Likert scale appeared, and they were asked to rate how natural the target sentence is based on the context.

The manipulation of the context ([±con] and the manipulation of the intervening adverb ([high] vs. [low]) resulted in four total conditions. There were a total of four test trials and four filler trials. The conditions were organized in a Latin-square design for a within-subject study, so each participant saw one of the four conditions once.4

Acceptability judgments: All of the target sentences were rated quite low, the average per condition ranging between 2.5 and 3.4 out of 6. The average ratings per condition are shown in Table 1. To determine whether the average ratings across different conditions differed significantly from each other, we ran homoscedastic t-tests. In the high adverbial conditions, the average rating was significantly higher in a conservative context than in a non-conservative context (p = 0.004). The average ratings did not show any significant difference other than this.

Truth value judgments: Figure 4 plots the average proportion of ‘true’ responses by adverb type and construal. There was a main effect of construal, but no main effect of adverb type. For both high and low adverbials, participants were significantly more likely to accept the target sentence when the proportional measure was compatible with a conservative construal (76% and 66% vs. 28% and 23%; Z = 5.091, p < 0.001).

As in Study 1, it is not the case that all sixty-one participants reliably accepted or rejected the target sentences. However, we focus on statistically significant contrasts between the conditions to inform us of the general availability of the readings. We first examine the contrast in the rate of acceptance in the [low] conditions. What we see is a significant difference between the conservative and the non-conservative conditions. That is, the rate of accepting a target sentence containing an intervening low adverbial is much higher (66%) in a conservative context (66%) than in a non-conservative context (23%).

While we predicted both of the [high] conditions to be rated as true, this was only the case with the [high,+con] condition. What we see is that the non-conservative reading is not available even when a high adverbial interferes between the host noun and the quantifier. The exact nature of this low acceptability remains open. We suggest, however, that semantic restrictions for scrambling in Korean could provide a possible explanation for this. Since the high adverb is base-generated above the subject, the subject (the host noun) must undergo scrambling to the left of the high adverb to a position above TP in all our [+high] contexts. It is noteworthy that scrambled phrases tend to be interpreted definite/specific, which in turn suggests that the host noun would strongly favor a definite/specific reading in the [+high] conditions. In Study 1 which we presented in Section 3, we observed that definite host nouns are not compatible with a non-conservative reading of a measure phrase. Given that our high adverb conditions favor definite/specific nouns due to scrambling, our high adverb conditions are in fact biased towards a conservative reading, which may explain low acceptability of [+high,–con] contexts. It must be noted, however, that the same issue does not arise for low adverb conditions. The subject is base-generated higher than the low adverb within vP, so that intervention by low adverbs does not necessarily involve subject scrambling. Put differently, there is no semantic or syntactic bias to interpret the subject to be definite or specific in low adverb conditions. Thus, the validity of our claim based on low adverb intervention would not be affected by the issues concerning subject scrambling in Korean.

The average acceptance rates also depend on the specific adverb we used in each of the target sentences. For example, as Figure 5 shows, there is some variation across the different lexical items. As mentioned earlier in the paper, it is not clear whether the adverbial twinuckey (‘late’ or ‘belatedly’) we categorized under low adverbials is in fact a low adverbial that is merged within vP and linearized together with the subject when vP undergoes Spell-Out. We see that both conservative and non-conservative conditions involving twinuckey show low acceptance rates, unlike the other three low adverbs we tested. It is unclear at this point what this tell us about the nature of the adverb.

Taken together, our results support the adnominal approaches of proportional floating quantifiers: the non-conservative construal does not readily arise when the low adverb intervenes, suggesting that the proportional floating quantifier is base-generated in an adnominal position.

Our paper provides empirical evidence for adnominal quantifiers that result in a non-conservative reading on the surface. In other words, we show that adnominal quantifiers can result in non-conservative readings when moved out of the DP. This analysis can be extended to taypwupwun (‘most’) in Korean, which has previously been analyzed as an adverbial quantifier in Park 2007. The predictions of an adnominal account for Study 1 and 2 hold for taypwupwun as well, according to the authors’ judgments. For example, when the host noun is definite-marked with a demonstrative and a plural marker as in (36), the non-conservative reading is not available.

Also, when a low adverbial intervenes taypwupwun and its host noun, the non-conservative reading is unavailable as shown in (37).

Thus, we can uniformly analyze taypwupwun and proportional quantifiers in Korean as adnominal floating quantifiers.

We have seen in Study 1 that the floating proportional quantifier is ambiguous between a non-conservative and a conservative reading. A question remains on why such an ambiguity holds for floating proportional quantifiers in Korean. There are at least three possible explanations for this. First, because the value of the restriction is determined by focus, it is possible that the participants interpreted the target sentences as involving a focused VP. This is discussed in both Herburger 1997 and Ahn & Sauerland 2017 as a possible derivation of a conservative reading. The nature of the online study involving text stimuli does not allow us to control for this or determine whether this was indeed the interpretation made by participants who accepted the conservative reading. Second, it is possible that the participant interpreted the bare floating quantifier to be underlyingly case-marked, thus equating (38) with (39). Case-marked quantifiers are argued to carry a covert pro, thus necessarily resulting in a conservative reading, as the gloss in (39) shows (Ko 2014).

Lastly, Ko (2014) claims that bare numeral quantifiers, which are normally adnominal in neutral contexts, can be interpreted as adverbials when they receive an exhaustive focus. Specifically, Ko argues that when an exhaustive focus falls on a bare numeral, the numeral is merged on a focus projection separate from the host noun and is not subject to the EG, patterning like other adverbial floating quantifiers as shown in (40)–(42). The analysis for exhaustive-marked floating quantifiers also involves a covert pro, thus resulting in a conservative reading.

This potential ambiguity of bare numerals explains rather unexpected judgment variations concerning the EG in some focal contexts, and can be extended to account for the ambiguity observed in our studies. In particular, it is possible that the participants were interpreting the target sentences as carrying exhaustive focus on the quantifier.

We have argued in this paper that the non-conservative reading resulting from a floating quantifier is adnominal. Adnominal quantifiers that result in a non-conservative reading are not so common across languages, and require a very specific syntactic structure and focus in many languages as shown in Herburger 1997 and Ahn & Sauerland 2017. A question that is raised immediately is why this type of quantifier is so rare in natural languages. One possibility is that this is due to semantic and syntactic economy constraints. An event modification through an adverbial quantifier usually results in a truth-conditionally identical proposition, and is often syntactically simpler. Adnominal FQs necessarily involve movement of the host noun, which can be restricted by various conditions in syntax. By contrast, adverbial FQs can be freely merged in the VP domain as long as its semantic type can be matched with the verbal domain. Adjunction is a free operation in syntax whereas movement requires morphological or featural triggers to be licensed. This might explain why, even in languages that allow adnominal floating quantifiers, stranding of an adnominal quantifier is rare compared to that of an adverbial quantifier (Ko 2014).

The results of our studies have implications on non-conservative quantifiers identified in previous literature. While Herburger (1997) and Ahn & Sauerland (2017) both discuss the possibility of deriving a non-conservative reading out of a quantifier through movement and association with focus, there has not been any overt, empirical evidence for this. Our study confirms that the floating quantifier structure involved in non-conservative construals in Korean are indeed derived through a movement from the nominal domain, thus forming the first set of empirical evidence for the adnominal accounts of non-conservative quantifiers.

While we have combined the discussion of the two adnominal accounts into one category in this study, the two analyses do have nontrivial differences. For example, Herburger (1997) crucially relies on the non-conservative quantifiers involving D-raising rather than QR, while Ahn & Sauerland (2017) argue that non-conservative proportional quantifiers QR to take clausal scope. It is an open question whether the two kinds of movement can be teased apart using data from Korean. This is left for future investigation.